Source-linked AI summary
Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine
Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, Renqian Luo, Scott Mayer McKinney, Robert Osazuwa Ness, Hoifung Poon, Tao Qin, Naoto Usuyama, Chris White, Eric Horvitz
TL;DR
It remains unclear whether modern generalist foundation models can match specialist medical capabilities without domain-specific training or fine-tuning. This paper systematically explores general-purpose prompting for GPT-4, introducing Medprompt, which tops existing medical benchmark results and generalizes beyond medicine.
Problem
It remains unclear whether modern large-scale foundation models can achieve exceptional medical performance without specialized training or fine-tuning, despite earlier advantages for domain-specific models.
Method
The study systematically explores inexpensive, general-purpose prompting strategies for GPT-4, combines them into Medprompt, and controls for overfitting with an eyes-off evaluation set.
Results
Medprompt tops existing results across standard medical question-answering datasets and surpasses 90% accuracy on MedQA for the first time.
Takeaways & Limitations
Prompt innovation can unlock specialist capabilities in a generalist model, with the strategy also studied across competency evaluations outside medicine.
Takeaways & Limitations
Strong benchmark performance cannot be taken to demonstrate real-world efficacy on open-world healthcare tasks.
Abstract
from arXiv · showhide
Generalist foundation models such as GPT-4 have displayed surprising capabilities in a wide variety of domains and tasks. Yet, there is a prevalent assumption that they cannot match specialist capabilities of fine-tuned models. For example, most explorations to date on medical competency benchmarks have leveraged domain-specific training, as exemplified by efforts on BioGPT and Med-PaLM. We build on a prior study of GPT-4's capabilities on medical challenge benchmarks in the absence of special training. Rather than using simple prompting to highlight the model's out-of-the-box capabilities, we perform a systematic exploration of prompt engineering. We find that prompting innovation can unlock deeper specialist capabilities and show that GPT-4 easily tops prior leading results for medical benchmarks. The prompting methods we explore are general purpose, and make no specific use of domain expertise, removing the need for expert-curated content. Our experimental design carefully controls for overfitting during the prompt engineering process. We introduce Medprompt, based on a composition of several prompting strategies. With Medprompt, GPT-4 achieves state-of-the-art results on all nine of the benchmark datasets in the MultiMedQA suite. The method outperforms leading specialist models such as Med-PaLM 2 by a significant margin with an order of magnitude fewer calls to the model. Steering GPT-4 with Medprompt achieves a 27% reduction in error rate on the MedQA dataset over the best methods to date achieved with specialist models and surpasses a score of 90% for the first time. Beyond medical problems, we show the power of Medprompt to generalize to other domains and provide evidence for the broad applicability of the approach via studies of the strategy on exams in electrical engineering, machine learning, philosophy, accounting, law, nursing, and clinical psychology.
1 Introduction
The paper asks whether modern generalist foundation models can achieve specialist medical performance without extensive domain-specific training. It introduces Medprompt, a systematically engineered prompting strategy that unlocks strong GPT-4 performance across medical challenge benchmarks.
- Modern foundation models demonstrate broad problem-solving capabilities, but their specialist performance without domain-specific training remains uncertain.
- Medprompt systematically steers GPT-4 with innovative, inexpensive, general-purpose prompting strategies rather than expert-crafted medical content.The design includes controls against overfitting during prompt engineering.
- Medprompt achieves state-of-the-art performance across standard medical question-answering benchmarks and substantially exceeds simple prompting and specialist models such as Med-PaLM 2.
- 9 absolute points: Medprompt’s accuracy gain on MedQA surpasses 90% for the first time on that benchmark.
- Ablation results show synergistic effects from combining in-context learning and chain-of-thought, with GPT-4 benefiting from authoring its own reasoning prompts.
2 Background
The background reviews why generalist foundation models may challenge domain-specific biomedical tuning and introduces prompting strategies used to steer them. It emphasizes that prompting can change task performance without updating model parameters, while ensembling increases computational cost.
- Earlier biomedical models benefited from domain-specific pretraining, but the relative advantage of that approach for larger modern foundation models is uncertain.
- Prompt engineering develops effective model inputs, and task performance can vary substantially without changing learned parameters.
- In-context learning solves new tasks from a few demonstrations without parameter updates and can produce effects similar to fine-tuning.
- Chain-of-thought prompting inserts intermediate reasoning steps before answers, and this paper investigates generating such demonstrations automatically with GPT-4.
- Ensembling combines multiple model outputs through methods such as consensus or majority vote, but increases inference requirements.Med-PaLM 2’s Ensemble Refinement used as many as 44 inferences for one question.
3 Experimental Design
The experimental design evaluates GPT-4 with prompting on medical benchmark datasets while explicitly addressing overfitting from iterative prompt selection. It also tests whether the approach generalizes beyond medicine.
- 3.1 Datasets: The primary evaluation uses 9 multiple-choice biomedical datasets from the MultiMedQA suite.
- 3.1 Datasets: MedQA uses 1273 four-option USMLE-style questions from the United States subset for comparison with prior work.
- 3.1 Datasets: MedMCQA evaluation uses its dev subset, containing 4183 questions with four multiple-choice answers.
- 3.1 Datasets: PubMedQA evaluation uses the reasoning-required setting, where models answer biomedical research questions using context from abstracts.
- 3.1 Datasets: The study evaluates medically relevant MMLU tasks spanning clinical knowledge, medical genetics, anatomy, professional medicine, biology, and medicine.
- 3.1 Datasets: Generality is tested on two nursing datasets and six additional MMLU datasets covering fields such as accounting and law.
- 3.2 Testing Methodology: 20%: The researchers randomly reserve this share of each benchmark dataset as an eyes-off evaluation set to reduce prompt-selection overfitting.
4 Power of Prompting: Exploration and Results
Medprompt composes dynamic few-shot selection, self-generated chain-of-thought, and choice-shuffle ensembling into a general-purpose prompting strategy. The approach systematically steers GPT-4 toward stronger medical benchmark performance while reducing reliance on expert-crafted prompts.
- 4.1 Dynamic Few-shot: Dynamic few-shot examples retrieve training cases similar to each test question instead of using one fixed exemplar set.The method uses similarity to identify relevant examples for individual task inputs.
- 4.2 Self-Generated Chain of Thought: GPT-4 generates chain-of-thought demonstrations from question–answer pairs, producing detailed explanations for few-shot prompting.The generated rationale is retained only when GPT-4’s predicted answer matches the ground-truth label.
- 4.3 Choice Shuffling Ensemble: Choice-shuffle self-consistency reduces position bias by repeating inference with reordered answer choices and aggregating the resulting answers.The ensemble uses the most frequent answer across the shuffled orderings.
- 4.4 Putting it all together: Medprompt: Medprompt combines intelligent exemplar selection, self-generated reasoning, and majority voting into a general-purpose prompt-engineering strategy.The integrated strategy is illustrated with additive component contributions on MedQA.
- 4.4 Putting it all together: Medprompt: The pipeline preprocesses training questions with embeddings and GPT-4-generated reasoning, retrieves similar examples, then ensembles predictions at inference.At inference, retrieved examples and their reasoning chains form context before repeated answer generation.
- 4.4 Putting it all together: Medprompt: 5 few-shot exemplars and 5 parallel API calls provide the reported cost–accuracy configuration for Medprompt.The authors describe this setting as balancing inference cost and accuracy.
- 4.4 Putting it all together: Medprompt: +0.4% performance on MedQA yields a 90.6% state-of-the-art threshold when using 20 exemplars and 11 ensemble items.The ablation suggests that increasing these hyperparameters can further improve performance.
5 Results
Medprompt achieves state-of-the-art performance across all nine MultiMedQA benchmark datasets, with eyes-off results indicating limited overfitting. Ablations identify chain-of-thought as the largest incremental contributor, while broader-domain tests support general-purpose applicability.
- GPT-4 with Medprompt outperforms all other models on every MultiMedQA benchmark.
- 5.1 Performance on Eyes-Off Data: 90.6% on eyes-on data versus 91.3% on eyes-off data suggests prompt engineering likely did not overfit MultiMedQA.Eyes-off performance was higher on 6/9 benchmark datasets.
- 5.2 Insights about Medprompt Components via Ablation Studies: +3.4% from chain-of-thought was the largest incremental MedQA gain, followed by +2.2% each from few-shot prompting and choice shuffle ensembling.The ablation used a 5 kNN-curated few-shot strategy and five ensemble API calls as its base Medprompt configuration.
- 5.2 Insights about Medprompt Components via Ablation Studies: 90.6% was reached with up to 20 few-shot exemplars and 11 ensemble steps, compared with 90.2% for standard Medprompt.The additional performance came with increased inference-time cost and complexity.
- 5.2 Insights about Medprompt Components via Ablation Studies: GPT-4 self-generated chain-of-thought outperformed the expert-crafted Med-PaLM 2 prompt by 3.1 absolute points on MedQA.Both prompts used fixed five-shot examples and no ensemble.
- Medprompt improved over zero-shot prompting by +7.3% on six diverse out-of-domain MMLU datasets, versus +7.1% on MultiMedQA.The out-of-domain comparison used zero-shot and five-shot baselines.
6 Limitations and Risks
The paper identifies benchmark contamination, limited real-world coverage, hallucinations, and bias as important limitations and risks of evaluating and deploying foundation models in healthcare.
- Benchmark contamination: Benchmark performance may reflect memorization or data leakage, despite blackbox testing and overfitting controls.The authors state that blackbox testing cannot guarantee that test data were unseen, and benchmark contamination concerns remain.
- Real-world validity: Medprompt’s benchmark results do not demonstrate efficacy on open-world healthcare tasks.Medical challenge problems use selected multiple-choice questions and do not capture the range and complexity of clinical practice.
- Reliability: Hallucinations may compromise model generations and advice, and improved prompting could make remaining errors harder to detect.The paper points to probabilistic calibration and trustworthy confidence measures as promising directions.
- Bias and fairness: Bias in foundation-model outputs remains a concern because optimizing overall performance may affect equitable performance across subpopulations.The authors emphasize balancing accuracy with equity to avoid exacerbating healthcare disparities.
7 Summary and Conclusions
The study shows that systematic prompt engineering can steer GPT-4 to strong specialist performance on medical challenge benchmarks without special fine-tuning or expert-authored prompts. Medprompt tops existing medical benchmark results, generalizes across other fields, and should be explored alongside fine-tuning in high-stakes domains.
- Summary and Conclusions: Prompting steered GPT-4 to top-performing specialist capabilities without special fine-tuning or human specialist expertise for prompt construction.The evaluation included an eyes-off dataset and systematic exploration and combination of prompting strategies.
- Summary and Conclusions: Medprompt topped existing results across standard medical question-answering datasets, including results from Med-PaLM 2.Med-PaLM 2 used specialist medical data, fine-tuning, and handcrafted prompts from expert clinicians.
- Summary and Conclusions: GPT-4-generated custom chain-of-thought prompts outperformed hand-crafted expert prompts, while ablations assessed the relative importance of Medprompt components.The study also set aside eyes-off evaluation libraries to control for overfitting and examined performance in electrical engineering, machine learning, philosophy, accounting, law, nursing, and clinical psychology.
- Summary and Conclusions: Fine-tuning and other parametric updates may offer synergistic benefits, so both approaches should be explored in high-stakes healthcare domains.The authors frame this as a research direction rather than a demonstrated result.